Introduction

G protein-coupled receptors (GPCRs) are important pharmaceutical targets comprising the single largest structural family of gene products in the human genome and are characterized by seven transmembrane helices (TMs) [1]. Various computational approaches have been applied to the discovery of new GPCR ligands [2,3,4,5]. Structure-based approaches have sampled chemical space broadly to reveal new chemotypes as agonists or antagonists of various GPCRs, which may then be optimized structurally. Computational approaches for GPCR ligand discovery may be generally more efficient than high throughput screening of assembled chemical libraries. Another productive approach is to use GPCR structure-based or homology models and docking/molecular dynamics to guide the modification of known ligands by rational design [6]. Recently, machine learning (ML) has become a promising tool in medicinal chemistry for systematic drug discovery, in general, and with respect to GPCRs specifically [5]. ML techniques harness the power of algorithms to analyze vast structure activity relationship datasets, recognize patterns, and make predictions based on learned molecular features. ML has been applied to rapidly screen chemical databases, predict molecular interactions, and identify potential ligands [7,8,9] and can be used literally for end-to-end across drug discovery [7]. But the full potential of ML for GPCR ligand discovery has yet to be demonstrated [10].

In this study, we have focused our efforts on the modulators of purinergic signaling, an extensive signaling system relevant to many pathological conditions and the focus of drug discovery efforts. There are 19 cell-surface receptors in total in the signaling “purinome,” including 12 GPCRs (8 P2Y and 4 adenosine receptors). Among the P2Y purinergic receptors that have attracted interest for future therapeutics [11], the Gq protein-coupled P2Y6 receptor (P2Y6R) has emerged as an intriguing target due to its role in diverse physiological processes, including cell proliferation, inflammation, cerebroprotection, and immune responses [12, 13]. Activation of the P2Y6R has been implicated in various pathological conditions, including Alzheimer’s disease, Parkinson’s disease, epilepsy, pulmonary inflammation and fibrosis, diabetes, cardiovascular disease, and chronic neuropathic pain [12, 14,15,16,17,18,19,20,21]. Thus, it is an attractive target for the development of antagonists as novel pharmaceutical agents. However, there is no P2Y6R structure available, and there are currently few selective P2Y6R antagonists (Fig. 1). In contrast, the structure activity relationship (SAR) of P2Y6R agonists, including those with high affinity, has been reported [22,23,24,25]. The most frequently used P2Y6R antagonist in pharmacological studies is MRS2578, which is an irreversibly binding diisothiocyanate derivative [26, 27]. Chromene derivatives have also been explored as P2Y6R antagonists but are of only moderate affinity [28, 29]. Recently, Zhu et al. [30] reported apparently competitive P2Y6R antagonists in the class of 2-(1-(tert-butyl)-5-(furan-2-yl)-1H-pyrazol-3-yl)-1H-benzo[d]imidazole derivatives. Representative X-ray crystallographic [31] and cryo-electron microscopic (cryo-EM) [32] structures of the P2Y1R, a member of the same Gq-coupled P2YR subfamily (P2Y1-like), are available and have been used as a template for modeling other related P2YRs [30, 33, 34]. In this study, we evaluated the interaction of the ML-identified compounds at three P2YR subtypes: the uracil nucleotide-preferring P2Y6R and P2Y14R and the adenine nucleotide-preferring P2Y1R. All three subtypes recognize nucleoside 5′-diphosphates as endogenous ligands and are related to inflammatory pathways [11]. Thus, antagonists could have translational potential.

Fig. 1
figure 1

Known ligands of the P2Y6R with approximate affinity values. A Agonists. B Antagonists. See references [11, 22,23,24,25,26, 28,29,30] for details

Full size image

Materials and methods

Data curation and machine learning modeling

Public data available on P2Y6R agonists in ChEMBL [35] was found at ChEMBL4714 which was curated and used to build ML models. Collaborations Pharmaceuticals’ proprietary software “E-Clean” was used to “clean” and average activities for datasets prior to model building in Assay Central (AC) [36]. “E-Clean” handles duplicate compounds by either averaging, removing, or keeping duplicates based on InChIKey. For these data, duplicate molecules with continuous activity data were first converted to -logM and then were averaged. “E-Clean” logs the SMILES strings of the duplicate compounds along with their activities and indices for inspection by the user. If needed, compounds were also subjected to charge neutralization, salt removal, and standardization via custom software using open-source RDKit functions. The standardization within AC was done as follows: A simple standardization workflow consisting of the following steps and using the Indigo Toolkit was applied: read molecule from the string representation (e.g., SMILES or MOL), generate InChI and InChIKey, use InChIKey to find and remove duplicates, dearomatize, remove enhanced stereo, remove unknown stereo, standardize and reposition, if necessary, stereo bonds (e.g., wedged bonds), standardize or flag erroneous charges, flag erroneous valences, remove isotopes, remove dative and hydrogen bonds, remove smaller component if multicomponent chemical, flag multicomponent chemicals, neutralize. All chemicals which are duplicates or flagged with errors (e.g., erroneous valences or charges) are then excluded from the result, but all erroneous or duplicate records are included into a protocol associated with a given dataset and available for review in the user interface. The proprietary AC software uses the P2Y6R agonist (EC50) data (244 molecules) at either a cutoff of 1 mM or 5 mM with multiple algorithms integrated into the web-based software to build classification models. The algorithms include: deep learning (DL), adaboost classifier (ada), Bernoulli NB (bnb), k-nearest neighbors (kNN) classifier, logistic regression (lreg), random forest classifier (rf), support vector classification (SVC), and XGBoost (XGB) classifier models with Extended Connectivity Fingerprint (ECFP)6 descriptors. In all cases, fivefold cross validation was performed except for deep learning for which we removed 20% of the training set, in a stratified manner for the classification models, and these were used as external test sets for models trained on the remainder of the data.

Pharmacological assays

Hit compounds for pharmacological screening were purchased from MedChemExpress (MCE, Monmouth Junction, NJ, USA) and Millipore Sigma (St. Louis, MO, USA). Stock solutions (5 mM, DMSO) of the non-nucleotide test compounds were stored at − 20 °C. Selective P2Y1R agonist MRS2365 ([[(1R,2R,3S,4R,5S)-4-[6-amino-2-(methylthio)-9H-purin-9-yl]-2,3-dihydroxybicyclo-[3.1.0]hex-1-yl]methyl] diphosphoric acid mono ester trisodium salt) was from Tocris (Minneapolis, MN). UDP was from Millipore Sigma (St. Louis, MO, USA).

Calcium mobilization assay

In order to identify potential agonists or antagonists for human P2Y6R or P2Y1R, hit compounds from our ML models were tested using the FLIPR assay with Calcium 6 dye kit (Molecular Devices, CA) in 1321N1 astrocytoma cells either with stably expressing hP2Y6R or hP2Y1R [22, 31]. Briefly, 1321N1-hP2Y6R or -hP2Y1R cells were grown in a 96-well black plate (2 × 104cells/well) for 24 h. Cells were treated with different concentrations of antagonist in presence of calcium 6 dye for 45 min, and assays were performed with a FLIPR-Tetra System (Molecular Devices, CA). Ester-protected dye is absorbed into the cytoplasm during incubation, is cleaved, and binds to calcium. Intracellular calcium is released upon P2Y6R activation with UDP (100 nM final concentration), or P2Y1R activation with selective agonist MRS2365, and interacts with the dye which was monitored using a FLIPR-Tetra. For agonist screening, cells were incubated with dye for 45 min followed by addition of hit compounds which were diluted in 1 × Hanks balanced salt solution (HBSS) buffer with 20 mM HEPES buffer at fixed concentration (80 µM). The IC50 values for different antagonists or % of activation at 80 µM of agonist were determined using a three-parameter logistic equation in GraphPad Prism software (Version 10.1.1, GraphPad, San Diego, CA). The results are presented as mean ± SEM (n = 2–3), unless noted with each molecule [29].

Competitive binding assay

Hit compounds identified using our ML models were tested in CHO-hP2Y14R cells [37] by fluorescent-based competitive binding assay. CHO cells stably expressing human P2Y14R were grown in 96-well plate and when cells were 80% confluent incubated with different hit compounds with a single concentration (400 µM) for 30 min at 37 °C in 5% CO2 incubator. Fluorescent antagonist MRS 4174 (20 nM) [37] was added, and incubation continued for another 30 min. Cells were washed thrice with DPBS and detached with Cellstripper (Corning, Glendale, AZ, USA) followed by resuspension in DPBS. Acquire the mean fluorescent intensity (MFI) using flowcytometry (CytoFLEX, Beckman Coulter, Brea, CA, USA), and determine the percentage of inhibition. The mean autofluorescence of cells was measured in the absence of the fluorescent ligand. The mean fluorescence intensity in the presence of fluorescent ligand was corrected by subtracting the autofluorescence. Data analysis was performed with GraphPad Prism software (Version 10.1.1, GraphPad, San Diego, CA, USA) and presented as mean ± SEM (n = 2–3) [37].

Results

Selection of compounds

We used the public data available in the CHEMBL database for the P2Y6R agonists (EC50) (ChEMBL4714) to build classification machine learning models using the algorithms: DL, ada, bnb, kNN, lreg, rf, SVC, and XGB with our AC software at different cutoffs (Table 1). Models showed generally good fivefold cross validation statistics and we selected a model built with a cut-off at 5 μM to score the following compound libraries: Microsource (2560 compounds), CNS-Penetrant compound library (MCE, 833 compounds), clinical compound library (MCE, 1977 compounds), and compounds from our internal projects (> 200 molecules) using consensus predictions. We identified a small set of 19 candidate molecules for screening at P2Y6R (structures shown in Fig. 2). The structures include mostly known experimental and approved drugs, including anti-infective compounds, anticancer agents, an anti-asthmatic drug, an antipsychotic drug, proton pump inhibitors (PPIs), a dietary supplement, and several others available to us. Four of the antiviral agents have a uracil nucleoside-related structure. Three additional molecules were selected without ML.

Table 1 Machine learning classification model 5-fold cross validation statistics for P2Y6 agonists (EC50): (A) cut-off of 1 µM (83 actives, 161 inactives) and (B) cut-off of 5 µM (139 actives, 105 inactives)
Full size table
Fig. 2
figure 2

Structures of diverse compounds chosen for testing as ligands of the P2Y6R and other P2YRs. Compounds were selected using ML models with the exception of dexlansoprazole, mivebresib (ABBV-075), and INCB-057643, which were selected by similarity to other compounds tested

Full size image

Pharmacological evaluation

A total of 22 compounds (Fig. 2) were assembled for initial functional screening at the human (h) P2Y6R expressed in 1321N1 astrocytoma cells. Compounds were selected using ML models with the exception of dexlanzoprazole, mivebresib, and INCB-057643. The proton pump inhibitor (PPI) dexlansoprazole, which also has anti-fibrotic activity [38], was tested because its racemic form lansoprazole showed some activity initially. Mivebresib, a pan inhibitor of the bromodomain and extraterminal (BET) family of bromodomains [39], and INCB-057643 [40] were tested because they are both BET inhibitors and showed Tanimoto similarity (MACCS fingerprints) > 0.60 compared to anticancer drug ABBV-744 (0.61 for INCB-057643 and 0.65 for Mivebresib). The latter two compounds were selected after we discovered the P2Y-related activity of ABBV-744, a 6-methyl-7-oxo-6,7-dihydro-1H-pyrrolo[2,3-c]pyridine derivative that is a selective inhibitor of the BD2 domain of BET family [41]. Mivabresib and INCB-057643 have the same core heterocyclic structure as ABBV-744. In addition to anti-cancer activity, ABBV-744 also impedes SARS-CoV-2 infection by regulating the host response [42].

The assay consisted of measuring the ability of each compound to inhibit calcium transients in the cell induced by the native P2Y6R agonist, UDP (200 nM, Table 2). The initial screen was at a single concentration (400 µM), which was set relatively high to lower the bar for detecting positive hits. Compounds that inhibited by > 50% at that concentration were run in full concentration–response curves to obtain an IC50 value. Two compounds (hit rate ~ 20%) displayed the most potent inhibition, thus warranting the determination. ABBV-744 and AZD5423 (having a 1-phenyl-1H-indazole scaffold) were found to have IC50 values of 75.7 µM and 574 µM, respectively (Fig. 3). Thus, ABBV-744 was the most interesting among the tested compounds as a putative P2Y6R antagonist. This compound is an experimental myelofibrosis and cancer drug [43] (clinicaltrials.gov, NCT04454658, accessed July 21, 2023) that acts as an inhibitor of BET bromodomain proteins, specifically BD2 domain of BRD2, BRD3 and BRD4. AZD5423 is an inhalable non-steroidal glucocorticoid receptor modulator that is in clinical trials for mild allergic asthma and COPD [44] (clinicaltrials.gov, NCT01226316, accessed July 21, 2023). There is no apparent structural similarity between the two uncharged P2Y6R antagonistic hit compounds, ABBV-744 and AZD5423.

Table 2 Assay of selected ML compounds at the hP2Y6R and the hP2Y14R
Full size table
Fig. 3
figure 3

Concentration-dependent inhibition of the hP2Y6R, in the presence of a fixed concentration of agonist (UDP, 100 nM), by two hit compounds, ABBV-744 and AZD5423, measured using a FLIPR to detect calcium transients in stably transfected hP2Y6R-expressing 1321N1 astrocytoma cells (representative curves shown). The IC50 values determined are listed in Table 1 and in the text

Full size image

The same compounds were tested in P2Y6R agonist mode, i.e., for the ability to stimulate calcium transients at 80 µM in the same stably transfected cell line in the absence of UDP. However, none of the compounds displayed significant, potential P2Y6R agonist activity. ABBV-744 and AZD5423 stimulated calcium transients to only 8.3% and 24%, respectively, relative to the full agonist (UDP, 200 nM) set as 100%. None of the compounds exceeded 26% increase of calcium transients at 80 µM. Therefore, ABBV-744 and AZD5423 are not partial agonists at the P2Y6R.

The effects of increasing fixed concentrations of these two putative antagonists, ABBV-744 and AZD5423, on the concentration-dependent P2Y6R activation by UDP are shown in Fig. 4. The antagonism by both compounds appears to be insurmountable, suggesting that they are not acting as competitive P2Y6R antagonists.

Fig. 4
figure 4

Concentration-dependent activation of the hP2Y6R by UDP and its inhibition by two hit compounds, ABBV-744 and AZD5423, measured using a FLIPR to detect calcium transients in stably transfected hP2Y6R-expressing 1321N1 astrocytoma cells (representative curves shown). AZD3965 failed to inhibit the UDP effect in this full curve experiment

Full size image

In a previous study by Puhl et al. of ML-selected ligands of adenosine receptors (ARs) [45], we found unanticipated interactions with other AR subtypes than the originally targeted A1AR subtype. Thus, we considered that there might be some overlap of activity at other P2YRs, in a similar fashion. The compounds were therefore examined in a binding assay at the Gi-coupled P2Y14R, which is similar to the P2Y6R in that both are activated by uracil nucleotides, including UDP, although P2Y14R is Gi-coupled in the P2Y12R-like subfamily of P2YRs. The binding assay, which we developed and have used extensively to screen for P2Y14R antagonists, is based on the inhibition of binding of a selective, high affinity fluorescent ligand MRS4174, containing AlexaFluor488 [37]. This ligand is used in a whole cell assay (stably transfected hP2Y14R-expressing CHO cells) in which competitive binding was measured by flow cytometry. Although two compounds were not included in the P2Y14R binding assay, none of those tested potently inhibited P2Y14R binding at a concentration of 100 µM. Experimental ophthalmic drug TAK-593 and experimental Aurora B/C kinase inhibitor GSK1070916 [46, 47] registered only 50% and 38% inhibition, respectively, of P2Y14R fluorescent binding at this concentration, and all of the other compounds tested produced < 20% inhibition. The inhibition by ABBV-744 was only 15%. We chose not to elevate the primary screening concentration to 400 µM, because of the previous observation of interference in the fluorescent binding with various compounds at > 100 µM concentrations.

Finally, the hit compounds were tested in a functional assay of P2Y1R activity (Fig. 5), as indicated by calcium transients in stably transfected hP2Y1R-expressing 1321N1 astrocytoma cells. Unexpectedly, ABBV-744 at 30 µM modestly enhanced activation of the P2Y1R by selective agonist MRS2365 (Fig. 5A). ABBV-744 at 100 µM produced a more robust enhancement of the P2Y1R activity (Fig. 5B). ABBV-075 at 30 µM appears to have a slight P2Y1R agonist activity (Fig. 5C).

Fig. 5
figure 5

Comparison of functional effects (in a FLIPR assay of calcium transients) at the hP2Y1R stably expressed in 1321N1 astrocytoma cells. A, B Effects of ABBV-744 and several other ML-selected compounds on concentration-dependent hP2Y1R activation by selective nucleotide agonist MRS2365: A 30 µM ABBV-744, AZD3695, and AZD5423; B 100 µM ABBV-744, AZD3695, and AZD5423. C Effects of ABBV-744 and ABBV-075 on the activation of the hP2Y1R

Full size image

Conclusion

The principal hit compound, experimental anticancer drug ABBV-744, an epigenetic reader domain inhibitor, had multifaceted interactions with the P2YR family. It inhibits hP2Y6R activation by UDP in a non-surmountable fashion, suggesting that it is not a competitive antagonist based on Ca2+ mobilization, but additional studies of different signaling pathways will be needed. The precise mechanism of inhibition was not determined in this study and will be explored in later experiments. The same compound enhanced hP2Y1R activation by MRS2365, a selective agonist, but lacked orthosteric binding affinity at the hP2Y14R. Other ML-selected compounds were either weak (another anticancer drug, AZD5423) or inactive in inhibiting the hP2Y6R. We did not discover any novel hP2Y6R agonists, which was the initial ML strategy. Weakly inhibiting compounds at the hP2Y14R were TAK-593 and GSK1070916. Nevertheless, as in our previous study of ML for identifying adenosine receptors ligands [45], activity at closely related subtypes of the same GPCR family, or other atypical activities at the targeted subtype, seems to occur more often than not. Thus, we have identified new leads for using small molecules to modulate the P2Y6R as well as other P2YRs. The multifaceted activities of ABBV-744 need to be directly compared to other P2YRs not studied here, as well as other purinergic signaling proteins such as P2XRs. A ML approach such as that demonstrated has the potential to enable repurposing of approved or experimental drugs based on previously undetected activities.